【作者】 李粉兰；

【导师】 徐可欣；

【作者基本信息】 天津大学 ， 测试计量技术及仪器， 2005， 博士

【摘要】 自动人脸识别是一个典型的图像模式分析、理解与分类的问题,涉及到模式识别、图像处理、计算机视觉、智能人机交互、认知科学等多个学科,在公共安全、信息安全等领域具有潜在的应用前景。本文是针对开发一个鲁棒实用、用户规模庞大的人脸识别门禁系统所需要的核心技术和关键问题的解决方案而展开研究的,主要研究内容和创新点包括:1.研究了基于多个分类器级联的人脸检测方法和连续自适应MeanShift人脸跟踪技术,以实现系统中的实时人脸定位。为了对分割的人脸图像进行几何归一化,提出了一种基于支持向量机和Hough变换的眼睛自动定位算法。同现有方法相比,该方法无需设定参数,也无需建立模板,整个定位过程简单易行,对人脸在图像平面内的旋转、尺度及面部表情等变化也都具有一定的鲁棒性。从定位结果上来看,该方法的定位速度不仅较快,定位准确率也较高;2.研究了基于核函数直接判别式分析的人脸识别技术,以在一定程度上解决因光照、姿态等因素的适量变化产生的识别率下降的问题,并在该技术中首次采用了修正模型和分数次幂多项式核函数。通过在多姿态、多光照条件的人脸库上进行实验,结果表明该技术相对于现有其它的基于核函数的分析方法,具有更高的识别率,对上述因素变化具有更强的鲁棒性。通过在该技术中采用修正模型和分数次幂多项式核函数,进一步增强了该技术对上述因素变化的鲁棒性;3.为了降低基于Gabor变换的人脸局部特征的维数,提出了将模拟退火概念应用于传统遗传算法的混合遗传算法以用于最佳特征的选择。它不仅有效解决了传统遗传算法存在的“早熟”和优化停滞问题,而且相对于传统遗传算法具有更佳的优化性能和稳定性。为了与混合遗传算法形成对比,研究了基于余量的迭代搜索法和共轭梯度法。实验结果表明当采用共轭梯度法和混合遗传算法对特征进行选择时,尽管只选用了不到一半的特征,但识别率却分别提高了2.5%和6.25%,且在选用相同数目特征的情况下,混合遗传算法的识别率明显高于共轭梯度法的识别率,即通过特征选择不仅可以节省1倍以上的特征存储空间,而且识别率也可以得到一定程度上的提高;4.为了解决系统存在的光照问题,研究了光照锥方法和常用的图像处理光照补偿方法并进行了实验对比。实验结果表明相对于图像预处理方法,光照锥方法更能有效地消除图像中的光照变化对识别结果的影响。另外对如何选择用于构建锥体的训练图像进行了实验研究,结果表明选用光照方向变化适中的更多还原

【Abstract】 Automatic face recognition is a typical pattern analysis, understanding and classification problem, and is closely related to pattern recognition, image processing, computer vision, intelligent human-computer interaction, cognitive psychology etc. It is believed that face recognition has a great deal of potential application in public security, information security and other fields.In this thesis, the core techniques and key issues are studied, aiming at developing a robust and practical gate-entry control system based on face recognition for large-scale users. The main content and innovations of the thesis involve:1. The multi-classifier cascade face detection method and continuous adaptive MeanShift face tracking technique are studied to attain real-time face detection in the system. A support vector machine and Hough transform based eye localization method is proposed for face image’s normalization. Compare to existing eye localization methods, there is no need to set parameters and build templates in the proposed method and the localization is simple and easy to implement. At the same time, it is of definite robustness to face rotation in image plane、scale and facial expression variation. In terms of localization results, the localization velocity of the proposed method is faster and the localization accuracy is higher;2. In order to solve recognition rate reduction resulting from illumination、pose etc moderate variations to some extent, the Kernel Direct Discriminant Analysis method is studied and a corrected model and a fractional power polynomial kernel function are first employed in the method. Through experiments on face databases with multi-poses and multi-illumination conditions, the results indicate that in contrast to other methods based on Kernel, the method is of higher recognition rate and more robustness to the above variations. When corrected model and fractional power polynomial kernel function are applied, its robustness is improved further;3. Hybrid Genetic Algorithms based method which introduces Simulated Annealing concept into traditional Genetic Algorithms is studied to reduce the dimension of local facial features extracted by Gabor transformation. It not only effectively solves the“premature”and optimization stagnancy problems existing in traditional Genetic Algorithms, its stability and optimization performance is also更多还原